People escalate against a competitor labelled human and hold back against one labelled an optimising machine

arXiv:2609.21439 · cs.HC, econ.GN, q-fin.EC · Submitted 2026-09-18 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Computer Science. The day in computer science: software, systems and theory on arXiv.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "People escalate against a competitor labelled human and hold back against one labelled an optimising machine".

Jane: Here's the gist of the paper. People escalate against a competitor labelled human and hold back against one labelled an optimising machine The information effect is not explained by articulated strategy, or individual differences, and is consistent with a competitive response engaged when a human is a live possibility

Paper summary: Tom: So we’ve talked through the paper "People escalate against a competitor labelled human and hold back against one labelled an optimising machine." The main thing is that the way you describe the opponent matters as much as who they actually are.

Jane: It really boils down to separating what's happening inside the system from what's happening because of the information we provide. We see this split between the opponent effect and the information effect throughout >Page two>.

Lu: The paper shows that if you change how a human plays, that’s an opponent effect, but if you just change the label alone, it’s an information effect >Page two>.

Meng: For us in production, this means our prompting or our system design choices are acting like that labeling—they're affecting the behavior of the user without necessarily changing the core algorithm.

Lalam: I think what this means for users is that if they know they’re up against a highly optimized system, their level of commitment changes in a predictable way >Page one> <ref:2609.21439#pg3>. It gives them something concrete to react to.

Tom: The authors clearly show how those two things—the system and the label—interact through the experimental design >Page three>. It’s not just one thing causing the change.

Jane: The paper suggests that organizations have a lot of control over that information environment, which is why understanding this distinction is important for decision-making >Page two>.

Lu: I think the big picture here is figuring out how these effects combine in more complex, real-world applications, rather than just testing them in a controlled auction setting >Page two>.

Conclusion: Tom: So we've looked at how labeling something as human versus an optimizing machine changes how people react to it in this paper, and now we’re going to talk about what that actually means for us, Tom and Jane.

Jane: The authors are looking at a system where people compete against two different things: one labeled human, and one labeled an optimizing machine.

Lu: It really boils down to how the label you give something affects the whole interaction, whether it's a person or an AI.

Meng: From my side, what I’m thinking is that if we label our system as purely optimizing, people might pull back more than if we just say it's human.

Lalam: I agree with Meng there. The paper suggests the reaction isn't just about the actual tech; it’s about the perception of what you’re facing.

Tom: Exactly. It's not some deep strategy thing, or even individual differences that explain this effect; it seems tied to how we frame the competitor.

Jane: So if you change the label, that changes how people engage with the system, and it’s not just a quirk of psychology.

Lu: The paper shows that this information effect is consistent with a competitive response engaged when a human is still a live possibility in the mix.

Meng: That suggests our labeling choices are part of our design, not just semantics on top of the code.

Lalam: It means if we're building something that needs to be deployed at scale, that label has real weight on how users behave and what they expect from us.

Tom: It makes you wonder how much control we actually have over these labels when you’re trying to get models running in the wild.

Jane: It really puts the focus on how we frame the interaction with our AI systems moving forward, which leads into thinking about the real-world deployment challenges.

Karlsruhe Institute of Technology (KIT) · Singularity AI

cs.HC, econ.GN, q-fin.EC

Submitted: 2026-09-18

Updated: 2026-11-10

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 89/100

The gist: People escalate against a competitor labelled human and hold back against one labelled an optimising machine The information effect is not explained by articulated strategy…

Key concepts

Human Competitor Label
This refers to a competitor explicitly identified as being human. The research shows that when facing this type of competitor, people tend to escalate their efforts in competition against them.
Optimising Machine Label
This label is applied to a competitor perceived as an optimising machine. When facing this type of opponent, people tend to hold back their competitive actions.
Information Effect
This phenomenon describes a change in behavior driven by the information available about the competitor's nature. In this case, the effect on behavior is not due to how strategically people choose to act or inherent differences between individuals.
Competitive Response
This is the overall pattern of action taken by people in response to a perceived competitor. The paper concludes that the observed difference in behavior is a competitive response specifically engaged when a human competitor remains a live possibility.

Terminology

Summary

People escalate against a competitor labelled human and hold back against one labelled an optimising machine The information effect is not explained by articulated strategy, or individual differences, and is consistent with a competitive response engaged when a human is a live possibility p. 0

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improve decision-making in competitive settings by adjusting disclosure based on opponent type. The information effect shows that being told that a human is a possible opponent raises prices while being told an optimising machine might be the opponent reduces prices.

  2. Design AI agents to exhibit behavior consistent with human responses when interacting with humans. This is achieved by training an agent using behaviour cloning and regularizing it toward the imitating agent so that it stays behaviourally plausible rather than drifting into inhuman play.

  3. Implement dynamic all-pay auction mechanisms where escalation is driven by incentives rather than external instructions. This structure ensures that escalation follows from the incentives rather than being staged.

  4. Develop AI systems capable of exhibiting different persistence levels based on their training objective when interacting with humans. The "optimising-machine label raises the round-to-round slope most (+0.203, p < 0.001), while the generic “AI label nearly as much (+0.196), while the imitating-AI label leaves it unchanged.

  5. Create decision support systems that recognize when opponent response speed affects behavior only against human-labeled opponents. The opponent effect of actual timing has a larger effect than any description, and "Responsiveness is higher among participants who in fact faced AI-Imitate (+0.047, p < 0.001) or AI-Optimal (+0.049, p = 0.001)."

Sources

Related papers